[Feature]: Ship assembly/HIP instead of binary #706
Description
Suggestion Description
Hi,
I notice that some aiter kernels are shipped directly as binary in https://github.com/ROCm/aiter/tree/main/hsa.
This is an issue for users who would like to modify the above kernels. For example, during AMD hackathon in Paris, we realized that a fused MOE assembly kernel shipped by aiter was bugged using TP=1 and Mixtral 7x8B model on MI300. Without a proper build system that allows to modify the kernels, it is very hard for contributors to attempt to fix the issues.
Ideally, these kernels should be shipped as HIP (or assembly if they are written directly in assembly), and aiter build system should generate the assembly on the fly when build aiter.
llvm-objdump is an option to disassemble, but it is not practical.
WDYT?
Operating System
No response
GPU
No response
ROCm Component
No response